xen: migrate vcpus only when needed
PATCH 1/1 xen: migrate vcpus only when needed
Currently, when pinning vcpus, the target vcpu is migrated to the first bit in the new cpumask even if the vcpu's current process is included in the new mask. This produces a non-obvious side-effect when pinning vcpus.
[x460-3 ~]# xm vcpu-list 5
Name ID VCPU CPU State Time(s) CPU Affinity
x460-3-guest 5 0 31 -b- 4.2 any cpu
[x460-3 ~]# xm vcpu-list 5
[x460-3 ~]# xm vcpu-pin 5 0 28-31
[x460-3 ~]# xm vcpu-list 5
Name ID VCPU CPU State Time(s) CPU Affinity
x460-3-guest 5 0 28 -b- 4.2 28-31
This patch adds a check to determine if a vcpu needs to switch processors during a pin operation, removing the above side-effect.
[x460-3 ~]# xm vcpu-list 2
Name ID VCPU CPU State Time(s) CPU Affinity
x460-3-guest 2 0 31 -b- 4.2 any cpu
[x460-3 ~]# xm vcpu-pin 2 0 28-31
Name ID VCPU CPU State Time(s) CPU Affinity
x460-3-guest 2 0 31 -b- 4.2 28-31
[x460-3 ~]# xm vcpu-pin 2 0 28-30
[x460-3 ~]# xm vcpu-list 2
Name ID VCPU CPU State Time(s) CPU Affinity
x460-3-guest 2 0 28 -b- 4.2 28-30
Tested on sedf scheduler.
Signature · 71 lines
Ryan Harper Software Engineer; Linux Technology Center IBM Corp., Austin, Tx (512) 838-9253 T/L: 678-9253 [email protected] diffstat output: sched_bvt.c | 14 ++++++++++---- sched_sedf.c | 12 +++++++++--- 2 files changed, 19 insertions(+), 7 deletions(-) --- # HG changeset patch # User rharper@x460-3 # Node ID ffa67b655a8f5855549880a0733c7fae6b86f3ae # Parent ad686e96536fe8c9318efa80e777d5c2555f2682 When setting a vcpu's affinity via pin operations, only migrate the vcpu if its current processor is not in the new affinity mask. Signed-off-by: Ryan Harper <[email protected]> diff -r ad686e96536f -r ffa67b655a8f xen/common/sched_bvt.c --- a/xen/common/sched_bvt.c Sat Apr 15 04:17:27 2006 +++ b/xen/common/sched_bvt.c Sat Apr 15 05:12:00 2006 @@ -287,11 +287,17 @@ if ( v == current ) return cpu_isset(v->processor, *affinity) ? 0 : -EBUSY; - vcpu_pause(v); + /* + * only migrate the vcpu if the vcpu's processor is not in + * the new cpumask_t + */ + if ( !cpu_isset(v->processor, *affinity) ) { + vcpu_pause(v); + v->processor = first_cpu(*affinity); + EBVT_INFO(v)->migrated = 1; + vcpu_unpause(v); + } v->cpu_affinity = *affinity; - v->processor = first_cpu(v->cpu_affinity); - EBVT_INFO(v)->migrated = 1; - vcpu_unpause(v); return 0; } diff -r ad686e96536f -r ffa67b655a8f xen/common/sched_sedf.c --- a/xen/common/sched_sedf.c Sat Apr 15 04:17:27 2006 +++ b/xen/common/sched_sedf.c Sat Apr 15 05:12:00 2006 @@ -1186,10 +1186,16 @@ if ( v == current ) return cpu_isset(v->processor, *affinity) ? 0 : -EBUSY; - vcpu_pause(v); + /* + * only migrate the vcpu if the vcpu's processor is not in + * the new cpumask_t + */ + if ( !cpu_isset(v->processor,*affinity) ) { + vcpu_pause(v); + v->processor = first_cpu(*affinity); + vcpu_unpause(v); + } v->cpu_affinity = *affinity; - v->processor = first_cpu(v->cpu_affinity); - vcpu_unpause(v); return 0; }
Quoted 4 lines
Currently, when pinning vcpus, the target vcpu is migrated to the first bit in the new cpumask even if the vcpu's current process is included in the new mask. This produces a non-obvious side-effect when pinning vcpus.
How is that non-obvious? You specify a set of CPUs on which you are happy for that VCPU to run and Xen may run the VCPU on any of those CPUs at any time. The current implementation's decision to change the CPU binding the moment you specify the set of CPUs, and always choose the lowest one, is perfectly reasonable given the specification of that interface.
-- Keir